Verify each finding against the current code and only fix it if needed.

Inline comments:
In @.github/workflows/build.yml:
- Around line 739-747: The upload step "Upload packaged app artifact" currently
uses actions/upload-artifact@... which only warns when globs match nothing;
update that step's inputs (the with: block for the actions/upload-artifact step)
to include the if-no-files-found option set to "error" so the job fails if the
Windows artifact globs (packages/desktop-electron/dist/*.exe, *.exe.blockmap,
latest*.yml) are missing; locate the step by its name or the
actions/upload-artifact usage and add if-no-files-found: error.

---

Nitpick comments:
In `@packages/opencode/test/github/build-workflow.test.ts`:
- Around line 117-123: The test currently extracts the selector step by position
using a small Ruby snippet that reads
data["jobs"]["select-build-target"]["steps"][0]["run"]; change that logic to
locate the step object whose id == "select" inside
data["jobs"]["select-build-target"]["steps"] and then read its "run" field
(i.e., iterate or find the step with step["id"] == "select" and return
step["run"]) so the helper is robust to inserted steps; update the Ruby snippet
(the variable ruby used with execFileSync) to perform this lookup and return the
run string.
- Around line 124-150: Replace the manual temp dir lifecycle (tmp created by
fs.mkdtempSync and later fs.rmSync) with the shared test helper tmpdir from
fixture/fixture.ts: create the temp dir via await using tmp = await
tmpdir("pawwork-build-target-") (or the helper's exact API), build outputPath
from that tmp, keep the spawnSync call and subsequent reading of outputPath
unchanged, and remove the explicit fs.rmSync cleanup so the tmpdir helper
automatically cleans up; if you cannot use await using, wrap the
mkdtempSync/spawnSync/read in a try/finally and move fs.rmSync into the finally
to guarantee cleanup.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In @.github/workflows/build.yml:
- Around line 740-745: The artifact upload step currently uses runner.os in the
job name and allows non-macOS runs but the path patterns under path only match
Windows (*.exe, *.exe.blockmap, latest*.yml), causing failures when Linux runs
occur; either restrict the step to Windows by changing the conditional to
runner.os == 'Windows' or make the artifact path conditional on runner.os (e.g.,
split into separate steps or use conditional path patterns) and keep the
if-no-files-found: error behavior; update the step that defines name,
if-no-files-found and path so the runner.os check aligns with the
Windows-specific patterns.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In @.github/workflows/build.yml:
- Around line 100-103: Update the x64 branch so the host uses an explicit Intel
macOS runner instead of macos-latest: change host="macos-latest" to
host="macos-15-intel" in the x64 case where platform_flag="--mac --x64" is set;
also update the corresponding test expectation in build-workflow.test.ts that
asserts the macOS runner for the x64 target to expect "macos-15-intel" now.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In @.github/workflows/build.yml:
- Around line 205-206: The workflow currently sets SELECTED_ARCH from
inputs.arch while SELECTED_TARGET already comes from
needs.select-build-target.outputs.target; to avoid drift, change SELECTED_ARCH
to use the selector output as the single source of truth (use
needs.select-build-target.outputs.target or a dedicated outputs.arch from the
select job) instead of inputs.arch so both SELECTED_TARGET and SELECTED_ARCH are
derived from the selector output (update SELECTED_ARCH assignment where it's
defined to reference the selector output symbol rather than inputs.arch).

In `@packages/opencode/test/github/build-workflow.test.ts`:
- Around line 12-49: The tests that spawn child processes and touch the
filesystem (those calling runSelectBuildTarget and the parseGithubOutput case,
e.g. the "selects the macOS x64 release matrix", "selects the Windows x64
release matrix", "rejects unsupported Windows release combinations", and
"ignores malformed GitHub output lines" tests) must be executed under the live
test harness; replace test(...) with the live variant provided by the file-level
testEffect harness (use it.live(...) or testEffect(...).live as used elsewhere)
so these cases run as live tests rather than normal unit tests; ensure each test
that calls runSelectBuildTarget or relies on real child-process/IO behavior is
converted to use the live wrapper.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In @.github/workflows/build.yml:
- Around line 69-71: Add `${{ inputs.channel }}` into the concurrency group
string in the workflow so the key becomes ${{ github.workflow }}-${{ github.ref
}}-${{ inputs.phase || 'submit' }}-${{ inputs.channel }}-${{ inputs.target ||
'macos' }}-${{ inputs.arch || 'arm64' }}, and update the corresponding
expectation in build-workflow.test.ts that asserts the concurrency group (the
test that checks the workflow concurrency string) to include the channel token
in the same position so tests match the new group format.

In `@packages/opencode/test/github/build-workflow.test.ts`:
- Around line 76-78: The test currently asserts the total count of
upload-artifact steps by checking uploadArtifactSteps.length (variable
uploadArtifactSteps created via steps.filter on step.uses), which is brittle;
instead locate the Windows-specific upload step(s) by filtering for the expected
step.name or other key fields (e.g., name === "Upload Windows artifact" or the
step.with.key fields) and assert those properties (name, uses, and the
with.path/with.name values) rather than asserting toHaveLength(3); update both
occurrences (around the uploadArtifactSteps filter and the similar block at
lines 118-123) to find and assert the named step(s) and their key fields.
- Around line 14-62: Several tests use Effect.promise(...) and Effect.sync(...)
directly; wrap their bodies in Effect.gen(function* () { ... }) instead. For
each test referencing runSelectBuildTarget or parseGithubOutput, replace the
Effect.promise(async () => { ... }) and Effect.sync(() => { ... }) callbacks
with Effect.gen(function* () { ... }) and move the existing async/sync logic
into that generator body so the test follows the repository's Effect test
pattern (keep the same assertions and calls to
runSelectBuildTarget/parseGithubOutput, just relocate them inside the
generator).